Enabling Performant and Flexible Model-Internal Observability for LLM Inference
Provides a practical systems primitive for model-internal observability in LLM inference, addressing the need for timely internal state access in inference-time workloads.
DMI-Lib enables low-overhead access to LLM internal states during inference, achieving 0.4%–6.8% overhead in offline batch inference and 6% average overhead in online serving, reducing latency overhead by 2x–15x over baselines.
Today's inference-time workloads increasingly depend on timely access to a model's internal states. We present DMI-Lib, a high-speed deep model inspector that treats internal observability as a first-class systems primitive, decoupling it from the inference hot path via an asynchronous observability substrate built from Ring^2, a GPU-CPU memory abstraction for capturing and staging tensors, and a policy-controlled host backend that exports them. DMI-Lib enables the placement of observation points across a rich space of internal signals and diverse inference backends while preserving serving optimizations and adhering to tight GPU memory budgets. Our experiments demonstrate that DMI-Lib incurs only 0.4%--6.8% overhead in offline batch inference and an average of 6% in moderate online serving, reducing latency overhead by 2x-15x compared to existing baselines with similar observability features. DMI-Lib is open-sourced at https://github.com/ProjectDMX/DMI.